Reflection’s 501B Beam and Utah’s First AI Doctor

Reflection AI launches Beam, a 501-billion-parameter open-weight model

The launch follows a funding round that valued Reflection at 25 billion dollars and major compute deals with SpaceX and Nebius for Nvidia GB300 chips. The company says it trained Beam Base in under four weeks, then ran a reinforcement learning phase on a cluster of around ten thousand GB300 GPUs that spun up roughly 1.3 billion training sandboxes in another four weeks. Sources: SiliconANGLE - AI, TechCrunch - AI, Latent Space

The claims have not been independently verified, and Reflection is benchmarking against GLM-5.2 even as newer Chinese models such as GLM-5.3, Kimi K3 and Qwen 3.8 Max are described in tier-3 coverage as generally ahead. Observers in that coverage place Beam around the GLM-5.2 level rather than as a clear win over the current leaders. Sources: SiliconANGLE - AI, TechCrunch - AI, Latent Space

Anthropic's charity machine builds ahead of its IPO

Documents Anthropic shared with potential investors show the company lets employees donate shares to charity and tops up those donations with additional company shares, with early staff receiving triple the value of their gift. Between October 2025 and March 2026 alone, that matching program cost more than 660 million dollars and diluted other investors' stakes. Source: The Decoder

Total donations in 2025 reached about 540 million dollars, nearly five times what Truist Financial and BlackRock gave that year at roughly 115 million and 109 million dollars respectively, putting Anthropic ahead of those Fortune 500 donors on those figures. CEO Dario Amodei and the six co-founders have pledged to give away at least 80 percent of their wealth, and employees influenced by Effective Altruism are already discussing where future gains should go, citing global poverty, AI safety and animal welfare. Source: The Decoder

The scale of giving is striking and it comes with a catch: the donations dilute other shareholders, and documents the company shared with potential investors suggest total giving could reach the billions after the IPO if the stock price keeps rising. Sources: The Decoder, Big Technology

OpenAI turns on invisible watermarks for ChatGPT in Europe

OpenAI is rolling out an invisible watermark called textGrain to ChatGPT and Codex users in the European Union over the coming weeks, citing the EU AI Act's requirement that AI-generated content be machine-readably labeled. The system nudges the model's word choices using a secret key so a detector can spot the pattern even after copy-paste, and in OpenAI's tests it detected watermarks in about 95 percent of 400-token passages about psychology at a one-percent false-positive rate. Sources: TechCrunch - AI, The Verge - AI, The Decoder

The company is not making watermarking a global default. API customers anywhere can opt in today for select models, support is coming to cloud partners like Azure, and detector access is limited to approved researchers and expert organizations. OpenAI stresses the system does not guarantee reliable detection and that a missing watermark does not prove human authorship. Sources: TechCrunch - AI, The Verge - AI, The Decoder

The approach gives OpenAI a workable compliance story for the EU, but OpenAI’s own data shows it is vulnerable to edits: replacing 10 percent of a passage's words can cut detection from about 92 percent to 66 percent, and swapping a quarter of the words drops it to 17 percent, making the watermark much easier to evade after modest editing. Sources: TechCrunch - AI, The Verge - AI, The Decoder

Altman says the world should accept 'some bad things' from AI

In an interview on Politico's Decoded podcast, Sam Altman said the world should accept some bad things happening for the benefits of AI, citing hacks and scams as a trade he would take because people will do orders of magnitude more good than bad with the technology. He framed OpenAI's stance as a lighter-touch middle ground between Anthropic's calls to slow down and those who want almost no rules, while warning that over-restricting AI could concentrate power in a few companies or one country. Sources: The Verge - AI, The Guardian - AI

The comments landed days after an OpenAI safety researcher resigned and called the company's culture broken. Critics including Florida governor Ron DeSantis and AI skeptic Gary Marcus pushed back hard on the idea that tech leaders should get to decide those safety tradeoffs for everyone else. Sources: The Verge - AI, The Guardian - AI

Amazon brings China's GLM 5.3 to Bedrock

Amazon has added Z.ai's GLM 5.3, a 753-billion-parameter mixture-of-experts model, to Amazon Bedrock, letting customers call it through managed APIs instead of running their own infrastructure. AWS pitches the model for coding and long-horizon agentic tasks and highlights its cybersecurity capabilities, demonstrating it running Strix, an open-source AI penetration-testing agent, in an authorized security test of a user's own application. Source: AWS machine learning blog

Bedrock layers on cross-Region inference and prompt caching, including an explicit caching mode for shared prompts that cuts latency and token costs when resending large chunks of context, and supports OpenAI-compatible APIs for calling the model. Source: AWS machine learning blog

Putting a model of this size behind managed endpoints with caching and familiar APIs gives enterprises access to a high-end coding and security model without having to operate that scale of infrastructure themselves. Source: AWS machine learning blog

Cohere rebuilds North into an enterprise 'control room' called North 2

Cohere has overhauled its agent platform into North 2, with a new orchestration system that lets agents handle multi-step tasks on their own, retain memory between sessions, and share reusable skills and knowledge libraries across a company instead of every team rebuilding the same workflows. From a chat interface, users can generate slide decks, dashboards or simple apps from a single text request, backed by templates and a drag-and-drop workflow builder. Sources: The Decoder, SiliconANGLE - AI

The new North Admin console tracks token use per user and per agent, lets administrators set quotas and company-wide caps with alerts, and adds security guardrails including PII screening, prompt-injection checks and autonomy policies that confine each agent to its allowed actions. The platform is model-agnostic and can run on-premises, in a virtual private cloud or fully air-gapped, with Bell Cyber and LG CNS already running it in production. Sources: The Decoder, SiliconANGLE - AI

Utah approves an AI app to prescribe acne treatment

Utah has become the first state to let an AI startup issue initial prescriptions, rather than just renewals. Nolla Health's pilot lets users talk to an AI agent, upload a photo of their face, and automatically receive a prescription for topical acne treatment without speaking to a doctor, drawn from a shortlist of physician-approved medicines. Sources: SiliconANGLE - AI, The Verge - AI

The rollout is phased: two physicians approve every AI-generated prescription for the first 100 patients, then the AI issues prescriptions directly while physicians review them afterward for the next 500 patients, and in the final stage physicians review at least 10 percent of prescriptions monthly. The program runs under Utah's AI Learning Lab sandbox and cannot prescribe isotretinoin or other oral medications. Sources: SiliconANGLE - AI, The Verge - AI